This article is naive to the point of being flat-out wrong, since it makes extremely naive assumptions about how a garbage collector works. This is basically another C++-centric programmer saying that smart pointers work better than the Boehm GC -- which is completely true but also completely misleading.
I'm not saying that GC is always the best choice, but this article gets the most important argument wrong:
> 1. Updating reference counts is quite expensive.
>
> No, it isn't. It's an atomic increment, perhaps with overflow checks for small integer widths. This is about as minimal as you can get short of nothing at all.
Yes, it is. Even an atomic increment is a write to memory. That is not "about as minimal as you can get short of nothing at all".
Additionally, every modern GC does generational collection, so for the vast majority of objects, the GC literally does "nothing at all". No matter how little work it does, a RC solution has to do O(garbage) work, while a copying GC can do O(not garbage) work.
Now, that's not to say that GC is automatically better. There are trade-offs here. It depends on the workload, the amount of garbage being created, and the ratio of read to write operations.
The article says:
> I've already stated I'm not going to do benchmarks. I am aware of two orgs who've already run extensive and far-reaching experiments on this: Apple, for use in their mobile phones, and the Python project.
I can counterpoint that anecdata: Google extensively uses Java in high-performance systems, and invented a new GC-only language (Go) as a replacement for (their uses of) Python.
The right answer is to do benchmarks. Or even better yet, don't worry about this and just write your code! Outside of a vanishingly small number of specialized use cases, by the time GC vs RC becomes relevant in any meaningful way to your performance, you've already succeeded, and now you're dealing with scaling effects.
I am finding this article and the comments here very surprising. Do a large number of people actually believe that the word "crypto" means exclusively "cryptocurrency"? Does anyone believe it means exclusively that?
To me it seems similar to how "auto" as a noun is generally short for "automobile", but most people are aware that other things can also be called "auto". When a camera says it is "auto focus" I cannot imagine that any normal person would assume that phrase has anything to do with automobiles.
It is incredibly common for the same word to have different meanings in different contexts. I personally have literally never had a conversation about cryptocurrency in which any person used the word "crypto" to mean "cryptocurrency", so I am clearly out of this loop. But if people decide to use it that way as slang in a certain context it certainly doesn't change the meaning of related words, or even mean it's impossible to use a different slang meaning in different contexts.
I'm not saying that GC is always the best choice, but this article gets the most important argument wrong:
> 1. Updating reference counts is quite expensive. > > No, it isn't. It's an atomic increment, perhaps with overflow checks for small integer widths. This is about as minimal as you can get short of nothing at all.
Yes, it is. Even an atomic increment is a write to memory. That is not "about as minimal as you can get short of nothing at all".
Additionally, every modern GC does generational collection, so for the vast majority of objects, the GC literally does "nothing at all". No matter how little work it does, a RC solution has to do O(garbage) work, while a copying GC can do O(not garbage) work.
Now, that's not to say that GC is automatically better. There are trade-offs here. It depends on the workload, the amount of garbage being created, and the ratio of read to write operations.
The article says:
> I've already stated I'm not going to do benchmarks. I am aware of two orgs who've already run extensive and far-reaching experiments on this: Apple, for use in their mobile phones, and the Python project.
I can counterpoint that anecdata: Google extensively uses Java in high-performance systems, and invented a new GC-only language (Go) as a replacement for (their uses of) Python.
The right answer is to do benchmarks. Or even better yet, don't worry about this and just write your code! Outside of a vanishingly small number of specialized use cases, by the time GC vs RC becomes relevant in any meaningful way to your performance, you've already succeeded, and now you're dealing with scaling effects.